Inversion of tunnel fires using limited monitored temperature data based on transfer learning approach and full-scale scenario applications

Numerical simulation coupled with deep learning models has proven effective for inversing fire source parameters in tunnel fires. However, models trained exclusively on numerical simulation data often struggle to adapt to diverse tunnel scenarios due to variations in tunnel geometry, sensor placement, and fluctuating heat release rate (HRR). To address these challenges, this study proposes a novel transfer learning framework. The model is pre-trained on numerical simulation datasets and uses temperature data from a limited number of sensors beneath the tunnel ceiling to inverse fire location and real-time HRR. The results demonstrate that the model achieves high accuracy using only six temperature sensors, even when positioned far from the fire source, achieving Coefficient of determination (R 2 ) values exceeding 0.99. Full model fine-tuning enhances the model's adaptability to variations in tunnel geometry, demonstrating remarkable performance in reduced-scale fire tests with R 2 values above 0.99 for fire location and 0.86 for HRR. The method is further validated in full-scale tunnel fires with highly variable HRR patterns. To handle sensor damage or data loss, the model utilizes temperature data from sensors farther from the fire source, maintaining R 2 values above 0.99 for fire location and 0.82 for HRR in the test set. Additionally, the model performs well in inversing HRR during the growth and stable periods of fires in both reduced-scale and full-scale tunnels, achieving Mean Absolute Percentage Error (MAPE) values below 0.2. This capability is critical for early fire detection and effective emergency response in real tunnel fire scenarios. • A transfer learning method proposed for tunnel fire inversion with sparse sensors. • Integrated fire test and simulation data to enhance cross-scale generalization. • Validated on full-scale tunnel fire tests under varied sensor availability. • Achieved fire inversion across growth, stable, and decay fire periods.